2026-08-15
Modeling AI as productivity shocks in a global CGE, enabled fossil emissions exceed avoided renewables emissions whenever fossil gains are nonzero; breaking even needs renewables gains 4-5x larger.
Most assessments of AI's climate impact draw the boundary around two things: the electricity datacenters consume, and the emissions AI applications can avoid, from grid optimization to end-use efficiency. The emissions-increasing side of the ledger, where AI lowers marginal costs in oil and gas exploration, drilling, and refining and makes previously uneconomic reserves viable, mostly sits outside the analysis. The IEA's 2025 Energy and AI report quantified a few of these effects, but its methodology is not public. This paper puts both directions into one global equilibrium model and computes the net.
The authors use GTAP-E-Power, a computable general equilibrium (CGE) model built on the GTAP-Power database (v11), aggregated to 20 sectors and 5 regions, with a 2017 base year adjusted so renewables generation approximates 2024 levels. AI is not modeled as an explicit sector; it enters as productivity shocks. The same capability applied upstream in oil and gas lowers extraction costs and expands economically viable supply; applied to solar and wind it raises output per installed unit. Because the model resolves cross-sector substitution, price feedbacks, and factor reallocation endogenously, it captures rebound and induction effects. Productivity gains on each pathway are calibrated to the empirical literature, with renewables deliberately set at the upper bound of documented technical potential (15-20% generation gains for mainstream technologies). Headline numbers use a parallel adoption benchmark: both pathways at the same adoption tier.
| Scenario | Annual CO₂ change |
| Parallel adoption (low-high tier) | net +0.47-1.8 Gt, 1.2-4.8% of 2024 global energy-related emissions |
| Fossil pathway only | +0.6-2.4 Gt, 3.3-13.3x the IEA's 2025 datacenter estimate (0.18 Gt) |
| High fuel-neutral (grid + demand-side) | only about -0.1 Gt |
| Carbon price $80/tCO₂ | net +0.7 Gt (enabled 1.0 vs avoided 0.3) |
| Carbon price $308/tCO₂ | net +0.1 Gt (0.3 vs 0.2) |
Breaking even requires renewables productivity gains 4-5x the fossil gains. Enabled emissions exceed avoided whenever fossil gains are nonzero. The direction holds across all 64 scenario combinations; varying two key elasticities by ±50% moves the enabled-to-avoided ratio between roughly 2-3x and 7x without reversing it. Decomposition locates the asymmetry upstream: extraction-only shocks roughly match combined shocks, while generation-only shocks slightly reduce emissions. Under parallel adoption, emissions grow faster than GDP, raising the carbon intensity of GDP.
For the AI industry's climate disclosures this is a direct challenge. Most "AI for climate" reporting lists compute footprint and avoided-emissions potential while leaving enabled pathways out of scope. This paper offers a reproducible quantification: the indirect effects exceed the datacenter footprint by an order of magnitude, and neither renewables acceleration nor carbon pricing alone closes the gap (at $308/tCO₂, parallel adoption is still net-positive). The policy implication is concrete: treat AI-enabled fossil productivity as a distinct supply-side lever to constrain. Teams building energy AI should also read the asymmetry as market intelligence: the empirical evidence for AI adoption across the oil and gas value chain is stronger than on the renewables side, which is itself a fact about where the money flows.
The authors list them at length: a static comparative framework with no temporal dynamics, infrastructure lock-in, or policy feedbacks; aggregation to 5 regions and 20 sectors obscures firm-level heterogeneity; physical constraints like grid physics and natural decline rates are abstracted away; green hydrogen, small modular reactors, and direct air capture have no sector, understating avoided potential. Two framing caveats sit on top: symmetric adoption and optimistic renewables parameters both push the headline numbers toward understatement, by the authors' own account. The flip side is that everything rests on how well the productivity shocks are empirically calibrated, and the fossil-side literature on AI adoption is itself thin. The direction is more trustworthy than the magnitude, which is how the authors present it.